Inequalities in Universal Health Coverage: Evidence from Vietnam

World Development Vol. 64, pp. 384–394, 2014 0305-750X/Ó 2014 Elsevier Ltd. All rights reserved. www.elsevier.com/locate/worlddev

http://dx.doi.org/10.1016/j.worlddev.2014.06.008

Inequalities in Universal Health Coverage: Evidence from Vietnam MICHAEL G. PALMER* The University of Melbourne, Carlton, Australia Summary. — Exploiting a window of opportunity in Vietnam, this paper examines the impact of social health insurance on target population groups. Significant inequalities in the coverage of service utilization and financial protection are found across groups. Persons with disabilities, and retirees to a lesser extent, experienced relatively high rates of service utilization and were most at risk of health care-induced poverty. A higher level of targeting in the design of benefit packages is recommended. Ó 2014 Elsevier Ltd. All rights reserved. Key words — universal health coverage, social health insurance, matching methods, South East Asia, Vietnam

1. INTRODUCTION

[for UHC] is not to cover everyone. Or even to give everyone the same cover. Rather, the challenge is really about narrowing inequalities in coverage” (Wagstaff, 2011). Remedying inequalities and injustices is a focus of the next development agenda in which no-one—irrespective of ethnicity, gender, geography, disability, and race—can be left behind and denied universal human rights and basic economic opportunities (United Nations, 2013). To this end, it is advocated that goals and social protection programs, including those around health, are designed to reach all population groups and excluded groups, in particular. This paper evaluates the impact of SHI across population groups in Vietnam. We exploit a window of opportunity in 2006 in which all SHI enrollees were eligible for the same benefits and co-payments. 1 Matching methods are applied to a rich collection of household and community living standards data collected in the Vietnam Household Living Standards Survey. The methodology matches insured persons with uninsured persons according to the target group characteristic and a range of other characteristics, such as age, level of education, employment sector, and geographical region. Providing that selection into insurance is determined by these characteristics, and that other distributional assumptions are met, then our results provide an unbiased estimate of the treatment effect (Abadie & Imbens, 2006). While the unconfoundessness principle cannot be directly tested, the Health Insurance Law (2008) in Vietnam identifies population target groups on the basis of defining characteristics and eligibility is mandatory for all groups except for farmers, the self-employed, and students for whom eligibility was group based at the time of survey. Unlike many other LMICs, SHI schemes in Vietnam are managed by a single administrative agency which further limits variability in the impact of SHI on population groups. Outcomes are evaluated against the objectives of UHC and include a range of utilization and economic burden outcomes. Applying a combination of matching estimators, we found significant inequalities in the use of health care and coverage of financial protection across target groups. Persons with

Universal health coverage (UHC) is widely recognized as essential to enhancing health, social cohesion and sustainable human and economic development (Evans, Marten, & Etienne, 2012; WHO, 2010). Guided by principles of providing access to all the services that people need without causing financial hardship, UHC is now a key policy goal of many low- and middle-income countries (LMICs). One significant means through which these countries aim to achieve UHC is the introduction of social health insurance (SHI) schemes (Giedion, Alfonso, & Diaz, 2013; Hsiao & Shaw, 2007; Lagomarsino, Garabrant, Adyas, Muga, & Otoo, 2012). Schemes often include a compulsory contributory scheme for civil servants and formal sector employees; a voluntary contributory scheme for the self-employed and employees in the informal sector; and a non-contributory membership for those that have limited capacity to pay a premium e.g., persons living below the official poverty line and other low-income and marginalized groups such as persons with disabilities, ethnic minority persons, and the elderly. Categorizing people in groups that represent common affiliations or identities is argued as an intrinsic aspect of human life, influencing individual well-being, capabilities, preferences, and behavior (Stewart, 2005). The impact of SHI is likely to differ across population groups due to both observable (e.g., health status, ability to pay, education) and unobservable (e.g., underlying health status, social capital) differences which jointly determine demand for healthcare (Zweifel & Manning, 2000). The supply of health services is further likely to differ across population groups. For example, the quality and accessibility of local health services will likely contrast sharply between formal employees and ethnic minority persons that live predominantly in urban and remote areas, respectively; and persons with disabilities encounter physical barriers which may limit their accessibility to the supply of health services. There currently exists little evidence on the efficacy of SHI in meeting the objectives of UHC in providing access to affordable care across different population groups. This can be explained by different administrative structures, eligibility conditions, benefits, and co-payments across SHI schemes. The lack of evidence on the inequalities in coverage represents a significant gap to inform the financing and targeting of SHI benefit packages. Benefits that reflect target populations’ needs are recognized as important to both the financial sustainability and equality of UHC (Giedion et al., 2013); “the challenge

* I am grateful to An-Ruo Bian for his assistance in the preparation of the results tables. I would also like to thank three anonymous reviewers, Jenny Williams, Sophie Mitra, Daniel Mont, and participants of the University of Melbourne Health Economics Group Annual Workshop and the International Health Economics Association 2013 Congress for their comments on earlier versions of this paper. Final revision accepted: June 4, 2014. 384

INEQUALITIES IN UNIVERSAL HEALTH COVERAGE: EVIDENCE FROM VIETNAM

disabilities, and the elderly to a lesser extent, experienced relatively high rates of service utilization and associated expenditures which induced a higher rate of poverty. The rest of the paper is organized as follows. The next section provides an overview on the development of SHI in Vietnam. This is followed by a description of the data and methods. The final section concludes with a discussion of the results, policy implications, and lessons learned for other LMICs on the path toward UHC. 2. BACKGROUND The story of the development of SHI in Vietnam is similar to that of many other LMICs. 2 A contributory compulsory scheme was first established in 1992 for public servants and employees in state-owned enterprises and the private formal sector in conjunction with a non-contributory scheme for social beneficiary groups including retirees, war veterans and their relatives (meritorious persons), and persons with disabilities. A voluntary scheme was subsequently introduced in 1994 for non-formal workers, especially farmers and the selfemployed, students and dependents of the compulsory scheme. In 2003, the non-contributory social beneficiary scheme was extended, under a funding arrangement known as Health Care Fund for the Poor (HCFP), to include households classified as poor, ethnic minorities in selected mountainous provinces, and households in especially socio-economic disadvantaged communes. From 2005, children under the age of six were added to the list of non-contributing groups. In 2008, the Health Insurance Law integrated existing schemes into one national program and identified 24 population groups that trace the development of the above schemes (Socialist Republic of Vietnam, 2008). The universality of a SHI system depends on its ability to enroll and collect premiums from the non-poor, and the government’s capacity to subsidize premiums for the poor or near-poor. Governments can compel formal sector employers to enroll their employees in SHI and deduct employee contributions from their salaries. Commonly, employers also pay a share of the SHI premium to create incentives for workers to enroll and avoid adverse selection. This is the case in Vietnam where the premium for formal employees is set at 6% of salary, with employees contributing 4% and employers contributing 2%. By contrast, enrollment of the self-employed and workers in the informal sector depends on voluntary contributions. Enrolling non-formal workers is a significant hurdle to universal coverage with premiums often subsidized by government according to an ability to pay schedule as is the case in Vietnam. 3 The premiums of the poor and nearpoor are usually fully or partially subsidized by government. In Vietnam, the rate is calculated at 3% of the minimum wage and is paid by the state. With formal employees typically comprising a small fraction of the LMIC population (10%) and low uptake of voluntary insurance among the non-formal sector, premiums received are often insufficient to cover benefits for the poor and nearpoor. Fiscally constrained LMICs may have to compromise and provide the poor and near-poor with fewer benefits than members of contributory and voluntary schemes. 4 This is the case for Columbia and Thailand, for example, where multi-tiered benefit systems exist (Hsiao & Shaw, 2007). 5 By contrast in Vietnam, the SHI benefit package is the same across all target groups and includes outpatient and inpatient treatment by public or registered private healthcare providers (however, the number of private providers remains low). 6 The

385

package is comprehensive and covers consultation, diagnostic tests, medical procedures and surgery, rehabilitation, and drugs. While the list of covered items is extensive, an expenditure cap of approximately US$35 per episode exists for hightech or high-cost services (Tran, Hoang, Mathauer, & Nguyen, 2011). A copayment of 20% was re-introduced in 2010 for all target groups except for meritorious 7 persons and children who were exempt. Retirees, the poor, and social beneficiaries incurred a copayment of 5%. While there is debate on the content of SHI benefit packages, it is generally agreed that services must be cost-effective and achieve both gains to health and protection against impoverishment from catastrophic medical expenses (Hsiao & Shaw, 2007). There is a growing literature on the impact of SHI in LMICs examining the impact of individual schemes due to different management, benefit, and eligibility arrangements. Typically, studies do not disaggregate by target group thus inequalities in coverage across the population remain largely unknown (e.g., Barros, Machado, & Sanz-de-Galdeano, 2008; Nguyen, 2012; Sepehri, Sarma, & Oguzoglu, 2011; Trujillo, Portillo, & Vernon, 2005; Wagstaff, 2010). Vietnam boasts a relatively high number of studies, partly due to the existence of high quality national survey panel data. Results are mixed across studies and schemes. For the HCFP, using 2004 and 2006 data Axelson, Bales, Pham, Ekman, and Gerdtham (2009) found a small positive impact on utilization and a strong negative impact on out-of-pocket expenditure whereas Wagstaff (2010) found no impact on use of services with an additional round of data and difference estimator. For the voluntary scheme, using the same two earlier rounds of data Nguyen (2012) reported a positive impact on inpatient and outpatient visits but no significant impact on out-ofpocket expenses whereas Jowett, Deolalikar, and Martinsson (2004) found a sizable reduction in expenses using self-collected cross-sectional data from three provinces. Sepehri et al. (2011) found no effect on outpatient expenditures for the compulsory and voluntary schemes and a modest expenditure reduction for the poor scheme from the 2004 and 2006 data. Vietnam presents a unique case study to evaluate universal health coverage as all SHI schemes are managed by a single agency (Vietnam Social Security) and benefit packages are uniform across groups. In the year 2006, copayments were also unified across groups and schemes hence our evaluation is confined to this period. Prior to 2005, formal employees and voluntary target groups incurred a 20% co-payment that was re-introduced in 2007 for voluntary target groups only. Another important consideration is that group eligibility requirements into the voluntary scheme (at least 10% of commune residents and school students participating) were still in place in 2006 which limits problems of adverse selection. 8 Furthermore, the 2006 Vietnam Household Living Standards Survey (VHLSS) is the only round of the survey to date which included a disability module, enabling identification of this important target group using a measure consistent with contemporary international classification of disability (Madans, Loeb, & Altman, 2010; Washington Group, 2008). In 2006, 52% of the Vietnamese population was covered by insurance with current estimates at approximately 60% (Somanathan, Huong, & Tran, 2013). Figure 1 presents the coverage of selected population groups identified in the Vietnam Health Insurance Law. High rates of coverage were recorded for state employees, the poor, and ethnic minority persons, and students at close to or above 80% coverage. Roughly half of non-state employees, the retired and disabled

386

WORLD DEVELOPMENT (%) 100 90 80 70 60 50 40 30 20 10 0

Figure 1. Insurance coverage by population group. Source: Author’s calculation based on 2006 Vietnam Household Living Standards Survey.

were insured with the lowest coverage rates recorded among the self-employed and farmers. 3. DATA The data in this study are drawn from the VHLSS 2006 (GSO, 2008). The survey contains extensive household and community information collected from 9,189 households (39,070 individuals) and 3,063 communes through a stratified cluster sampling design. The household questionnaire is categorized under broad domains of demographics, education, health, employment, income, assets, and expenditure. The health module is of particular importance to this study and includes information on self-reported health status, disability status, insurance status, and health care utilization. Health care utilization data include inpatient contacts in the previous 12 months, and outpatient and self-medication contacts in the previous 1 month, which were used to construct the individual health care utilization and expenditure outcome variables. The module further contained separate household-level information on health care expenditures in the last 12 months, which were used to construct the variables associated with the economic burden of health care. The commune questionnaire contained important community level variables that may determine the demand for health care, such as the poverty status and remoteness of the commune. 9 (a) Definition of population groups Eight target groups commonly identified in SHI systems were selected for analysis, and were defined as follows 10: (i) Formal employee: individuals reporting the following work sectors as the most time consuming job in the last 12 months: state-owned economic sector, collective economic sector, private economic sector or foreign-invested economic sector (n = 3,747). (ii) Retiree: individuals reporting the reason for not working in the last twelve months as retired/too old (n = 1,979). (iii) Person with disabilities: individuals reporting having “a lot” of difficulty or “cannot do” in at least one of the six functional domains: vision, hearing, concentrating,

walking, self-care, and communication (n = 1,265). Following (Mont & Cuong, 2011), this is a high (severe) threshold of disability and is consistent with insurance eligibility requirements. (iv) Ethnic minority person: individuals reporting an ethnicity, other than Kinh or Chinese, from 54 ethnic groups (n = 6,188). (v) Person living in conditions of poverty: individuals reporting as living in households classified as poor within their commune (n = 5,036). (vi) Farmer: individuals reporting their occupation as a skilled/unskilled worker in agriculture, forestry, or aquiculture that worked within the self-employed sector (n = 12,264). (vii) Self-employed person: individuals reporting as working in the self-employed sector that were not farmers (n = 6,523). (b) Measures of outcomes To estimate the impact on health care utilization, we examine contacts across the full range of service and facility types in Vietnam. As a meter for quality of care, we disaggregate inpatient and outpatient contacts across commune clinics, and public hospitals at district and provincial/central levels. 11 For inpatient contacts, we include only public facilities since the number of private facilities (hospitals or clinics) offering inpatient services in Vietnam was very few. For outpatient contacts, the number of private hospital visits was low and merged with private clinics to form a single category of private facilities. Unlike previous versions, the VHLSS 2006 collected information on the number of inpatient days so we use this variable as a proxy for usage intensity. To estimate the impact on financial protection, we examine several measures including health expenditures, catastrophic expenditures, and health payments-adjusted poverty. Unlike previous studies, we disaggregate health expenditures for formal services: consultation/treatment, on-site medication, tips, diagnostic tests, off-site medication, and travel. Catastrophic expenditures are defined as household health expenditure per capita exceeding a threshold budget share of household non-food expenditure per capita over the previous year. 12

INEQUALITIES IN UNIVERSAL HEALTH COVERAGE: EVIDENCE FROM VIETNAM

Three thresholds (10%, 20%, and 40%) are applied to provide a cross-section of incidence. Household income net of health expenditures per capita is measured against the official 2006 poverty line to provide a measure of health payments-adjusted poverty. 13 4. METHODS To estimate the impact of SHI on target population groups we apply propensity score matching (PSM), a popular method to minimize selection bias in cross-sectional non-experimental data under certain identifying assumptions (e.g., Heckman, Ichimura, & Todd, 1997; Jalan & Ravallion, 2003a, 2003b). The first is the conditional independence assumption (CIA) that assignment to treatment D is independent of potential outcomes Y, given a set of observable characteristics X which are not influenced by treatment: [Y(0), Y(1) \ D|X]. For the groups under study, selection into insurance on the basis of expected outcomes is likely to be contained. Selection into insurance was mandatory with the exception of farmers, the self-employed, and students for whom eligibility was group based at the time of survey. By law, all state employers and private companies of greater than 10 employees are required to register their employees in the SHI system. Upon retirement, formal employees are rolled over into the non-contributory retiree compulsory scheme. People with disabilities and beneficiaries of the HCFP are identified by local Commune Committees. While it is conceivable that informal sector workers or students could move into communes or schools that met the required threshold of enrollees to purchase voluntary insurance or workers could transfer into the formal sector to qualify for compulsory insurance, these scenarios are unlikely given the low-perceived benefits of insurance. It is more likely

387

that non-compliance of formal employees to compulsory enrollment or the eligibility criterion of poverty and disability status is influenced by unobservable factors. For example, a near-poor household with a history of bad health or a person with a non-severe disability but high health care needs be awarded a free health insurance card by Commune Committees. The second identifying assumption is the common support condition which requires that the probability (P) of assignment to treatment for persons with the same X values is positive between zero and one: [0 < P(D = 1|X) < 1]. This is to rule out perfect predictability of D and ensure that there are untreated matches for the treated for every x. Figure 2 displays the predicted probability or “propensity score” distributions of the treated (above the horizontal line) and the untreated (below the horizontal line) for each population group. The range of treated and untreated distributions is similar across all groups as indicated by the relatively low number of observations that are dropped (“off support” given by the lighter shade of gray). However, there remains an issue with distributional overlap, typically there are fewer untreated than treated observations with which to match in the higher end of the distribution. In this case, an untreated observation may have to be used multiple times and the estimated program impact will not be reliable in this range. This problem is particularly acute for formal employees, retirees, persons with disabilities, and farmers. Following Bales, Knowles, Axelson, Manh, Luong, & Oanh (2007), we apply a suitable algorithm that drops treated observations when the number of untreated observations was less than 5% of the number of treated observations (trimmed distributions are shown in Figure A1). 14 If the two identifying assumptions of CIA and common support hold, together known as strong ignorability, then population level average treatment effects can be estimated;

Figure 2. Untrimmed probability distributions of the treated and untreated by population group. Notes: Treated/untreated are above/below the horizontal line. Source: Author’s calculation based on 2006 Vietnam Household Living Standards Survey.

388

WORLD DEVELOPMENT

the most prominent of which is the average treatment effect on the treated (ATT). ATT is relevant to this study since it estimates the effect on persons for whom the program was targeted. It is defined in (1) as the difference between expected outcomes with and without treatment, conditional on X for those who participated in treatment, where D is a binary variable denoting treatment status and takes a value of 1 for the treated and 0 for the untreated. The counterfactual outcome for the treated can never be observed yet strong ignorability implies that allocation to treatment or nontreatment is random such that the mean outcome of the untreated individuals can serve as the counterfactual mean, as defined in (2). There remains the practical problem of conditioning on the vector X when the number of x is large, dubbed the “curse of dimensionality.” Rosenbaum and Rubin (1983) show that if potential outcomes are independent of treatment conditional on X they are also independent conditional on a so-called propensity score P, the conditional probability of receiving treatment given observed covariates X, as defined in (3). sATT ¼ EðsjD ¼ 1Þ ¼ E½Y ð1ÞjD ¼ 1; X   E½Y ð0ÞjD ¼ 1; X 

ð1Þ

sATT ¼ EðsjD ¼ 1Þ ¼ E½Y ð1ÞjD ¼ 1; X   E½Y ð0ÞjD ¼ 0; X 

ð2Þ

sATT ¼ EðsjD ¼ 1Þ ¼ E½Y ð1ÞjD ¼ 1; P ðX Þ  E½Y ð0ÞjD ¼ 0; P ðX Þ ð3Þ The PSM estimator for the ATT (3) is the mean difference in outcomes of the treated and untreated over the common support, appropriately weighted by the propensity score distribution of participants. Conditioning on observable characteristics over a common support and in the absence of any pre-treatment unobservable differences between the treated and untreated, PSM yields an unbiased estimate of ATT. (a) Propensity score estimation Propensity scores were estimated using a probit model. 15 The CIA implies that only variables that simultaneously influence participation in insurance and demand for health care be included in the model. Furthermore, variables must not be affected by participation in treatment or the anticipation of treatment. Variables were selected on the basis of our understanding of the institutional context and economic theory (Grossman, 2000; Zweifel & Manning, 2000). Target group identifiers play a particularly important role in participation and outcome hence we estimate separate propensity scores and matching procedures for each subpopulation (as defined in Section 3(a)). The same model specification is used for each target group since groups can be enrolled across a variety of SHI schemes and hence are subject to the spectrum of SHI eligibility criterion. Included variables which influence SHI eligibility are those defined in Section 3(a) plus variables defining the omitted target groups. A range of demographic, socio-economic, health, household, and community level variables were further included. Following advice in the literature, we include higher power variables on age and income (proxied by household consumption expenditure per capita) to best balance the distribution of variables (Caliendo & Kopeinig, 2008). Results from the probit models are presented in Table A3 and are generally consistent with factors that we would expect to be correlated with being insured.

(b) Matching estimators As defined, the PSM estimator for ATT is the difference in outcomes between treated and untreated individual/s over the common support, weighted by propensity scores. The question remains what weights to apply to match the treated with the untreated units. We tested five commonly used matching algorithms for each sample: nearest neighbor, radius, kernel, local linear, and Mahalanobis matching (Becker & Ichino, 2002). We briefly describe the three matching techniques and the testing criteria adopted in this paper as well as the choice of variance estimator (for a detailed review refer Guo and Fraser (2010)). Nearest neighbor matching (NN), as the name suggests, finds the untreated observation/s with the closest propensity score to the treated observation whereas radius matching (RM) matches all untreated units within a specified radius of the propensity score of the treated unit. We allow for multiple matching (up to 10 matches) for the NN estimator, and set a radius equal to one-quarter of the standard deviation of the propensity score on the treated sample for the RM estimator as recommended by Rosenbaum and Rubin (1985). In both instances, weighting is defined by the inverse of the number of untreated matched to the treated, and matching is applied with replacement since the number of untreated units is low at the high end of the propensity distribution in some samples. Kernel matching (KM) is a nonparametric matching estimator that uses weighted averages of all untreated units to construct the counterfactual outcome. The weighting function is the kernel density with weights assigned according to the distance of propensity scores to the treated unit. We use the default Epanechnikov kernel function and test for three bandwidths (0.04, 0.06, and 0.08) with higher bandwidths allocating higher weights over a wider range of propensity scores. To satisfy the CIA, the distribution of the covariates in the treated and non-treated groups must be balanced. We apply several procedures to assess the quality of matching estimators (Baser, 2006). In the initial model building stage we applied an algorithm developed by Dehejia and Wahba (2002). The algorithm stratifies the sample until no statistically significant difference exists between the mean propensity scores of the treatment and non-treated groups and then tests the difference in covariate means between the two groups in each stratum. Second, we assess the mean (and median) standardized bias defined as the difference in the covariate means between the treated and matched control groups as a percentage of the square root of the average of sample variances in both groups (Rosenbaum & Rubin, 1985). Third, we assess the reduction in the pseudo R2 of the probit model after matching. As shown in Table A1, standardized mean and median biases are less than the 5% rule of thumb threshold, and the pseudo-R2’s are one percent or less which indicates that observable characteristics explain very little of the variation in the propensity scores in the combined treated and matched comparison sample. Estimation of ATT variance is bootstrapped at 100 repetitions and adjusted for clustering. There remains some question over whether bootstrapping is valid for matching methods. Abadie and Imbens (2008) show that single NN matching estimators are not asymptotically efficient, and develop a robust variance estimation involving second-stage matching on a vector norm. While this is not established for the estimators used in this study, we apply the new variance estimator as a robust check. Specifically, we apply the bias-corrected estimator adjusted for heteroskedasticity, matching on four nearest neighbors following the recommendation in Abadie, Drukker, Herr, and Imbens (2004). 16

There is a positive impact of insurance on the probability of inpatient contact across all groups with greatest effect for the

(0.006) (0.004) (0.002) (0.006) (0.006) (0.002) 0.236 0.100 0.028 0.127 0.153 0.025 10,097 (0.008) (0.006) (0.003) (0.004) (0.005) (0.002) 0.261 0.118 0.038 0.052 0.072 0.019 6,513 (0.008) (0.005) (0.003) (0.008) (0.008) (0.003) 0.320 0.150 0.049 0.216 0.256 0.040 12,617 (0.016) (0.013) (0.009) (0.017) (0.017) (0.009) 0.365 0.206 0.083 0.456 0.535 0.079 5,031 (0.016) (0.010) (0.006) (0.018) (0.018) (0.006) 0.240 0.107 0.031 0.455 0.496 0.042 6,163 (0.017) (0.016) (0.012) (0.013) (0.014) (0.007) 0.497 0.297 0.122 0.155 0.207 0.052 1,264 (0.013) (0.011) (0.007) (0.008) (0.009) (0.005) (0.009) (0.007) (0.003) (0.003) (0.004) (0.002) 0.205 0.091 0.025 0.025 0.033 0.009 3,745

0.387 0.206 0.073 0.102 0.133 0.031 1,976

0.036 35.260 0.077 5.896 0.080 3.089 (0.003) (11.101) (0.005) (1.184) (0.005) (0.783) 0.048 103.289 0.109 13.181 0.139 9.620 (0.002) (5.443) (0.004) (1.447) (0.004) (0.705) 0.064 74.848 0.134 14.178 0.131 8.562 (0.005) (12.025) (0.007) (1.194) (0.007) (1.381) 0.082 99.680 0.138 9.686 0.118 7.646 (0.004) (8.209) (0.006) (1.460) (0.007) (0.542) 0.069 52.394 0.106 8.069 0.086 4.753 (0.012) (57.043) (0.015) (8.243) (0.014) (3.525) 0.186 401.203 0.322 51.255 0.279 34.495 (0.009) (31.172) (0.012) (6.101) (0.011) (7.538) 0.152 235.195 0.291 39.286 0.252 33.272 (0.004) (12.586) (0.006) (1.721) (0.006) (0.934) 0.066 92.732 0.098 11.471 0.092 7.911

(S.E.) Mean (S.E.) Mean (S.E.) Mean (S.E.) Mean (S.E.) Mean (S.E.) Mean (S.E.) Mean (S.E.) Mean

Self-employed Farmer Poor Ethnic Disabled

Table 1. Descriptive statistics of outcome variables by population group

Retired Formal

389

Source: Author’s calculation based on 2006 Vietnam Household Living Standards Survey.

(b) Estimation of treatment effects

Variables

There exist marked differences in demographic and socioeconomic characteristics across population groups (Table A2). As expected, retirees and people with disabilities were significantly older and students were significantly younger. Formal employees and the disabled were most likely to be male, and retirees and the poor were most likely to be female. Formal employees and informal workers were most likely to be married whereas the disabled and poor were least likely. One-quarter of retirees were disabled. Literacy rates and education achievements were highest among formal employees and lowest among people with disabilities, retirees, and ethnic minority persons. The majority of ethnic minority people and the poor worked as farmers, and approximately three-quarters of persons with disabilities were unemployed. Around two-thirds of formal employees worked in the nonstate sector. Incomes were highest for formal employees and retirees and lowest for the poor, ethnic minority persons, and farmers. Close to one-third of ethnic minority persons were officially poor, with relatively high rates of poverty recorded among the disabled and farmers. Farmers, ethnic minority, and poor persons were most likely to live in rural areas whereas formal employees, retirees, and the selfemployed were least likely. Over half of ethnic minority people lived in remote communes. Health care utilization similarly differs among subpopulations (Table 1). People with disabilities and retirees had significantly higher proportion reporting at least one inpatient visit in the previous year, at a rate more than double other insured groups. This was due to high hospital usage, particularly high level provincial and central hospitals (Table A4). Intensity of inpatient usage was also greatest for these two groups with the disabled and retired reporting double the number of inpatient days compared to other groups (Table A4). Inpatient expenditures were highest for people with disabilities, followed by retirees, and lowest for students and ethnic minority persons. Per inpatient visit, the disabled spent four times or more than most other population groups (and double the amount of retirees) due to relatively high treatment, medication, and travel costs (Table A4). Patterns in subpopulation utilization and expenditures were similar for outpatient and selftreatment services. Commune clinics were used more extensively than hospitals for outpatient care, especially among ethnic minority and poor persons, whereas district hospitals were the inpatient facility of choice for most population groups (Table A4). The economic burden of health care was greatest for the disabled with around 12% experiencing catastrophic expenditure at the highest threshold (40% non-food expenditure) compared to 3% for formal employees (Table 1). Retirees and the poor experienced similar, relatively high, levels of catastrophic health expenditures (7–8%) despite lower health spending among the poor. This is due to lower income and higher calculated poverty among the poor. After adjusting for health payments, the poverty rate among the poor increased by 8% representing the largest differential among population groups followed by the disabled with a 5% increase.

Student

(a) Descriptive statistics

Health care utilization Inpatient visit in last 12 months Inpatient expenditures per visit Outpatient visit in last month Outpatient expenditures per visit Self-treatment visit in last month Self-treatment expenditures per visit Economic burden of health care Catastrophic health expenditures 10% threshold 20% threshold 40% threshold Poverty Poverty net of health payments Poverty differential N

5. RESULTS

(0.002) (5.731) (0.003) (0.554) (0.004) (0.270)

INEQUALITIES IN UNIVERSAL HEALTH COVERAGE: EVIDENCE FROM VIETNAM

390

WORLD DEVELOPMENT

retired and disabled: a respective 13 and 9 percentage point increase between the insured and uninsured due largely to an increase in public hospital visits, particularly tertiary level hospitals (Table 2; Table A5). For all other groups, the impact of insurance on inpatient visits was modest at 1–4 percentage points and was mostly observed at district hospitals with the exception of ethnic minority persons who had higher commune clinic usage. Insurance increased the number of inpatient days for all groups with the highest increases among the disabled and retired at 2–3 days versus less than 1 day for other groups (Table A5). ATTs for inpatient expenditures were largely negative and statistically insignificant. Formal- and self-employees enjoyed lower on-site medication expenditures whereas retirees incurred higher off-site medication expenditures compared to their uninsured counterparts (Table A5). ATTs on total outpatient visits were mostly negative and statistically insignificant due to a substitution effect away from private facilities to public (particularly district hospital) facilities (Table 2; Table A5). The swing was strongest for formal employees, followed by the poor. By contrast, the net effect on total outpatient visits was positive for the disabled, retired, and self-employed due to relatively larger increase in district hospital usage. The impact of insurance on outpatient expenditures was largely negative with greatest effect for the retired and disabled due to lower medication and consultation expenditures, respectively. Insurance decreased the probability of a self-treatment visit for all groups by 4–7% with the exception of formal and informal employees for whom there was no impact (Table 2). There was no discernible impact on self-treatment expenditures.

Insurance consistently lowered the incidence of catastrophic health expenditures across groups at the lower threshold levels with little impact at the 40% threshold level (Table 2). Effects were greatest for retirees, formal employees, the poor and ethnic minorities with a 5–9% reduction at the 10% and 20% threshold levels. At the 40% level, formal employees had a 4% reduction in the rate of catastrophic expenditure. There was little impact of insurance on poverty with the exception of retirees and farmers who experienced an 8% and 5% increase in the rate of poverty, respectively. 17 Netted of health payments, poverty impacts were little changed as reflected by small and statistically insignificant poverty differential treatment effects. (c) Robustness checks Rather than matching on an estimated propensity score, we match directly on the covariates using a robust variance estimator developed by Abadie and Imbens (2002, 2006). In general, the ATT’s are slightly higher and statistically more significant than the PSM estimators (Table 3; Table A6). We summarize key differences relative to the PSM estimators. The inpatient per-visit expenditure effect for the disabled was approximately double, and off-site medication and travel expenditure effects associated with inpatient stays for the disabled were now statistically significant and higher than any other group. Inpatient expenditure treatment effects for the self-employed were reversed and statistically significant due to higher consultation and travel expenditures. Outpatient usage treatment effects were considerably higher for persons with disabilities and formal employees. For formal

Table 2. Propensity score matching treatment effects by population group

Health care utilization Inpatient visit in last 12 months Inpatient expenditures per visit Outpatient visit in last month Outpatient expenditures per visit Self-treatment visit in last month Self-treatment expenditures per visit Economic burden of health care Catastrophic health expenditures 10% threshold 20% threshold 40% threshold Poverty Poverty net of health payments Poverty differential

Formal

Retired

Disabled

Ethnic

Poor

Farmer

Self-employed

Student

0.007 (0.027) 54.872 (41.728) 0.010 (0.020) 3.057 (4.438) 0.005 (0.019) 1.235 (7.302)

0.126*** (0.013) 30.658 (81.744) 0.044 (0.027) 24.202** (11.467) 0.065** (0.034) 0.806 (19.700)

0.093*** (0.026) 103.062 (130.909) 0.034 (0.052) 20.351 (15.316) 0.047 (0.033) 2.088 (6.962)

0.015 (0.010) 39.589 (27.810) 0.029 (0.035) 1.671 (2.279) 0.067*** (0.025) 0.024 (0.857)

0.018 (0.014) 22.477 (24.553) 0.011 (0.017) 1.998 (2.820) 0.060*** (0.019) 0.313 (2.438)

0.028*** (0.011) 53.091 (49.073) 0.010 (0.014) 0.331 (4.273) 0.069*** (0.014) 0.456 (1.646)

0.015*** (0.006) 41.186 (27.149) 0.063*** (0.008) 5.951 (10.019) 0.003 (0.023) 4.912 (4.674)

0.003 (0.004) 14.229 (11.116) 0.001 (0.009) 1.577 (1.871) 0.037*** (0.011) 0.799** (0.332)

0.067** (0.032) 0.059*** (0.021) 0.035*** (0.013) 0.011* (0.006) 0.011 (0.007) 0.000 (0.006)

0.091*** (0.026) 0.080** (0.038) 0.056* (0.031) 0.076*** (0.024) 0.068*** (0.022) 0.008 (0.018)

0.052 (0.054) 0.090* (0.046) 0.017 (0.038) 0.009 (0.032) 0.021 (0.027) 0.012 (0.023)

0.064* (0.038) 0.051* (0.031) 0.020 (0.022) 0.018 (0.030) 0.017 (0.032) 0.001 (0.012)

0.083* (0.049) 0.046** (0.021) 0.009 (0.023) 0.006 (0.036) 0.002 (0.030) 0.004 (0.016)

0.046** (0.022) 0.043** (0.020) 0.017 (0.013) 0.045* (0.027) 0.049* (0.026) 0.004 (0.010)

0.039** (0.016) 0.032*** (0.009) 0.022** (0.010) 0.009 (0.012) 0.005 (0.013) 0.004 (0.005)

0.043* (0.017) 0.046*** (0.015) 0.005 (0.006) 0.010 (0.013) 0.014 (0.014) 0.004 (0.004)

Source: Author’s calculation based on 2006 Vietnam Household Living Standards Survey. *** Indicate significance at 1% level. ** Indicate significance at 5% level. * Indicate significance at 10% level.

INEQUALITIES IN UNIVERSAL HEALTH COVERAGE: EVIDENCE FROM VIETNAM

391

Table 3. Covariate matching treatment effects by population group

Health care utilization Inpatient visit in last 12 months Inpatient expenditures per visit Outpatient visit in last month Outpatient expenditures per visit Self-treatment visit in last month Self-treatment expenditures per visit Economic burden of health care Catastrophic health expenditures 10% threshold 20% threshold 40% threshold Poverty Poverty net of health payments Poverty differential

Formal

Retired

Disabled

Ethnic

Poor

Farmer

Self-employed

Student

0.0150*** (0.006) 42.489*** (15.026) 0.058*** (0.009) 8.391*** (2.513) 0.016* (0.009) 4.081*** (1.461)

0.118*** (0.018) 67.793 (90.318) 0.054** (0.022) 26.899** (12.434) 0.054** (0.023) 5.957 (10.421)

0.111*** (0.022) 213.435** (102.403) 0.107*** (0.025) 3.904 (11.287) 0.042* (0.024) 1.111 (5.395)

0.025*** (0.008) 8.349 (12.052) 0.042*** (0.011) 4.840*** (1.545) 0.063*** (0.008) 0.595 (0.623)

0.017** (0.008) 11.352 (19.872) 0.016 (0.012) 3.881*** (1.311) 0.055*** (0.013) 1.576 (2.178)

0.036*** (0.006) 32.097 (51.664) 0.004 (0.008) 2.003 (2.108) 0.028*** (0.007) 3.217** (1.645)

0.037*** (0.006) 46.521*** (15.832) 0.070*** (0.009) 2.847 (2.296) 0.014 (0.011) 1.176 (1.324)

0.013*** (0.004) 6.522 (10.640) 0.002 (0.005) 2.602*** (0.806) 0.027*** (0.006) 1.313*** (0.315)

0.054*** (0.018) 0.043*** (0.012) 0.017** (0.007) 0.019** (0.008) 0.016* (0.008) 0.004 (0.003)

0.075*** (0.024) 0.032* (0.019) 0.009 (0.013) 0.063*** (0.013) 0.080*** (0.014) 0.017** (0.007)

0.073*** (0.027) 0.066*** (0.024) 0.015 (0.018) 0.042** (0.018) 0.081*** (0.018) 0.039*** (0.009)

0.108*** (0.011) 0.044*** (0.006) 0.007*** (0.003) 0.026** (0.011) 0.044*** (0.009) 0.019** (0.008)

0.052*** (0.015) 0.014 (0.014) 0.014* (0.008) 0.019 (0.014) 0.003 (0.015) 0.022*** (0.008)

0.032*** (0.009) 0.009 (0.007) 0.008* (0.004) 0.009 (0.008) 0.016** (0.008) 0.007* (0.004)

0.019 (0.013) 0.003 (0.009) 0.012** (0.006) 0.015** (0.007) 0.019** (0.008) 0.004 (0.005)

0.028*** (0.009) 0.042*** (0.007) 0.014*** (0.004) 0.020*** (0.006) 0.020*** (0.006) 0.000 (0.003)

Source: Author’s calculation based on 2006 Vietnam Household Living Standards Survey. *** Indicate significance at 1% level. ** Indicate significance at 5% level. * Indicate significance at10% level.

employees this was due to a smaller substitution effect from private outpatient facilities whereas for the disabled a larger increase in the use of district hospitals and commune clinics. The poor recorded the highest substitution away from private outpatient facilities and the large effect for formal employees under the PSM estimator was all but removed and no longer statistically significant. Persons with disabilities experienced a 7% reduction in the rate of catastrophic health expenditures at the 10% and 20% threshold levels. Netted of health expenditures, the impact on poverty doubled for the disabled and was equivalent to the retired. Poverty increased by 4% for the disabled when taking account of health expenditures, representing the largest differential effect among population groups. An additional advantage of covariate matching is that we were able to estimate the average treatment effect for the control (ATC) group. If the ATT and ATC are similar in magnitude then it follows that the level of hidden selection bias and the departure from model assumptions is minimal (Guo & Fraser, 2010). We found moderate differences and in greatest effect for formal employees and ethnic minority persons.. As shown in Table A2, there exists considerable overlap among population groups which may influence our results. This is particularly the case for the retired and disabled with approximately one-quarter of the retired living with a disability and 40% of persons with disabilities are retired. Among the poor and ethnic minority persons, approximately 31% belong to the opposing group. To test the sensitivity of our results we excluded observations that were concurrently retired and disabled, and poor ethnic minority persons, for both propensity

score and covariate matching estimations. Utilization levels were stable whereas rates of health care-induced poverty were lower for the retired. Catastrophic spending and health care-induced poverty was slightly higher (lower) among the poor (ethnic minority persons). Overall, the results picture was little changed. To the extent that the trimmed samples differ to the treatment samples, our estimates may furthermore be biased. This risk is particularly acute for formal employees where 43% of the original sample was dropped. Covariate balancing between the treated dropped and treated sample for formal employees returned a mean bias of 24, consistent with other groups. We acknowledge that our method of including a large number of covariates in the propensity score estimation may exacerbate the common support problem and increase the variance of propensity score estimates, particularly in the smaller target group samples. We trimmed weakly correlated variables from the propensity score models yet found little variation in ATT results. Finally, it is possible that some variables are influenced by treatment, particularly the health status variables, and thus violate the CIA. Reduced form regression equations of health status variables did not reveal strong associations with insurance. 18 6. DISCUSSION AND CONCLUSION Exploiting a window of opportunity where benefits and co-payments were consistent across SHI schemes, this paper evaluates inequalities in UHC on target population groups

392

WORLD DEVELOPMENT

in Vietnam. In line with the principles of UHC, the study examines inequalities in service utilization and economic burden. A range of matching estimators was applied to account for observable differences between the treated and untreated, with unobservable differences contained by mandatory eligibility for all schemes with the exception of the voluntary scheme that was subject to group eligibility conditions. Robustness checks on the average treatment effect on the control groups suggest some degree of selection bias hence results must be viewed with caution. Results highlight strong inequalities in the impact of SHI across population groups. Insurance is found to increase the utilization of public health care, particularly inpatient care, across all groups. The effect is greatest for the retired and disabled due to high use of inpatient services at district and higher level provincial and central hospitals. The impact was relatively modest for other groups, especially ethnic minority persons, farmers and students, and was mostly observed at district hospitals with comparably little effect on commune clinic usage. Consistent with other studies (Axelson et al., 2009; Wagstaff, 2007), we find a strong substitution effect from private outpatient services to covered public outpatient services and in greatest effect for the poor. We also find strong negative effect on the use of self-treatment services across target groups. While these findings are positive from a public health point of view, they are contributing to well publicized problems associated with overcrowding of the Vietnamese public hospital system including long waiting times and poor quality consultations (e.g., Nhat, 2012). The findings point to a need for increased investment in the primary health care system of commune clinics and extending efforts to register private providers into the health insurance system. The impact of insurance on financial protection is less impressive. Directional effects on health expenditures were mostly negative yet statistically insignificant, consistent with other studies (Nguyen, 2012; Sepehri et al., 2011; Wagstaff, 2007). There is evidence of lower on-site medication expenditures associated with inpatient stays for formal employees and with outpatient visits of the retired. However, off-site inpatient medication and travel expenditures are higher for

the disabled, retired, and the poor. The impact is greatest for the disabled likely due to specialized medication and transport needs. Consistent with other studies (Axelson et al., 2009; Wagstaff, 2007), there is evidence that insurance reduces the incidence of catastrophic expenditures across target groups at the low and middle threshold levels, particularly for the retired, disabled, poor, and ethnic minority persons, with little impact at the high threshold level. The impact on poverty is mixed. There is a decrease in the incidence of poverty and poverty net of health expenditures for ethnic minority persons; the impact on both indicators is positive for the disabled and the retired. The impoverishing effect of out-of-pocket payment for healthcare is greatest for the disabled. Our findings, consistent with those of Wagstaff and vanDoorslaer (2003), suggest that it is non-hospital costs (medication and travel) that are pushing the insured into poverty. The modest impact of insurance on out-of-pocket payments may be partly explained by increased utilization of health care. In addition, the list of reimbursable items under insurance is not universal and expenditure caps exist for covered items. Hospitals are reported to routinely encounter drug shortages with insured patients purchasing medication from private pharmacies at their own expense (Somanathan et al., 2013; Tran et al., 2011). The above findings on the impact of insurance are a concern given the high average health expenditures and rate of catastrophic health expenditures across groups, particularly for the disabled and, to a lesser extent, the retired and the poor. They give support to calls to deepen coverage and reduce the economic burden on patients through increased Government subsidies or adjustments to the fee schedule for the insurance provider (Giang, 2011; Lieberman & Wagstaff, 2009; Xu, Evans, Carrin, Aguilar-Rivera, & Evans, 2007). Furthermore, benefit packages could be extended to include financially catastrophic treatments and medications. One consideration is to introduce a supplementary benefit package for the disabled and retired that include specialized services and assistive devices, medications, and travel entitlements.

NOTES 1. To the extent that members of population groups can be enrolled across various schemes with differing benefits or co-payments, it is not possible to evaluate the impact of insurance within population groups. For example, a person with disability may be enrolled in a non-contributory scheme for vulnerable groups or a contributory voluntary or compulsory scheme. 2. For a detailed summary of the development of SHI in Vietnam refer Ekman, Nguyen, Ha, and Axelson (2008). 3. Premiums range according to rural/urban and student status: VND 100,000 for rural students, VND 240,000 for rural residents, VND 120,000 for urban students, and VND 320,000 for urban residents.

6. In practice, the services an insured person actually receives may vary depending upon scheme enrollment and province of residence due to capitation-based payment schedules. However, the capitation system started with a small number of district hospitals in 2006 and thus is unlikely to be a major source of bias. 7. Persons and relatives determined to have made substantial contributions to the socialist revolution. 8. Group eligibility requirements to the voluntary scheme were dropped in 2008.

4. Furthermore, non-formal workers and the poor/near-poor may not receive the same benefits package as that financed by employer and employee contributions.

9. One complication was that the commune questionnaire was limited to rural communes. We cross-checked administrative records of urban communes categorized as poor and fortunately none of these communes were included in the VHLSS sample.

5. Refer Hsiao and Shaw (2007) and Giedion et al. (2013) for detailed description of individual country schemes.

10. Note target groups are not mutually exclusive in that respondents can be in more than one group.

INEQUALITIES IN UNIVERSAL HEALTH COVERAGE: EVIDENCE FROM VIETNAM 11. The number of visits at village health centers and regional clinics was relatively small and so were merged with commune health clinic visits. Similarly, central level hospital visits were merged with provincial hospital visits. 12. Non-food expenditure is recommended over total consumption expenditure to better reflect capacity to pay and distinguish between the rich and poor (O’Donnell, van-Doorslaer, Wagstaff, & Lindelow, 2008). 13. Note income is proxied by household consumption expenditure, which is considered a more permanent measure of well-being in LMICs (Deaton, 1997). 14. For each target group sample, observations with p score values greater than 0.6 was divided into 20 blocks of equal frequency and the minimum 5% criterion was applied (Bales et al. (2007). Trimming results are displayed in the Appendix.

393

15. The probit model better balanced the variables in a few of the target group samples compared with the logit model. 16. The bias-correction applies to all covariates which are the same set of covariates used in the propensity score estimators. We apply the default inverse of the sample variance matrix to measure the distance between the two vectors of covariates and elect the same number of matches (4) in the second stage matching to run the robust variance estimator. 17. The latter result was significant at the 10% level only. 18. All results are available from the authors upon request.

REFERENCES Abadie, A., & Imbens, G. (2002). Simple and bias-corrected matching estimators for average treatment effects. National Bureau of Economic Research Technical working paper series. Cambridge: National Bureau of Economic Research. Abadie, A., Drukker, D., Herr, J. L., & Imbens, G. W. (2004). Implementing matching estimators for average treatment effects in Stata. Stata Journal, 4, 290–311. Abadie, A., & Imbens, G. W. (2006). Large sample properties of matching estimators for average treatment effects. Econometrica, 74(1), 235–267. Abadie, A., & Imbens, G. (2008). On the failure of the bootstrap for matching methods. Econometrica, 76(6), 1537–1557. Axelson, H., Bales, S., Pham, M. D., Ekman, B., & Gerdtham, U. (2009). Health financing for the poor produces promising short-term effects on utilization and out-of-pocket expenditure: evidence from Vietnam. International Journal for Equity in Health, 8(20). Bales, S., Knowles, J., Axelson, H., Manh, P. D., Luong, D. H., Oanh, T. T. M. (2007). The early impact of Decision 139 in Vietnam: An application of propensity score matching. Report to the Ministry of Health, Health Strategy and Policy Unit. Hanoi. Barros, P. P., Machado, M. P., & Sanz-de-Galdeano, A. (2008). Moral hazard and the demand for health services: A matching estimator approach. Journal of Health Economics, 27, 1006–1025. Baser, O. (2006). Too much ado about propensity score models? Comparing methods of propensity score matching. Value in Health, 9(6), 377–385. Becker, S. O., & Ichino, A. (2002). Estimation of average treatment effects based on propensity scores. The Stata Journal, 2(4), 358–377. Caliendo, M., & Kopeinig, S. (2008). Some practical guidance for the implementation of propensity score matching. Journal of Economic Surveys, 22(1), 31–72. Deaton, A. (1997). The analysis of household surveys: A microeconomic approach to development policy. Dehejia, R. H., & Wahba, S. (2002). Propensity score matching methods for nonexperimental causal studies. Review of Economics and Statistics, 84(1), 151–161. Ekman, B., Nguyen, L. T., Ha, D. A., & Axelson, H. (2008). Health insurance reform in Vietnam: A review of recent developments and future challenges. Health Policy and Planning, 23, 252–263. Evans, D. B., Marten, R., & Etienne, C. (2012). Universal health coverage is a development issue. The Lancet, 380(8), 864–865. Giang, L. T. (2011). Reaching universal health insurance in Viet Nam: Challenges and the role of Government. Asia-Pacific Development Journal, 18(1), 47–72. Giedion, U., Alfonso, E. A., & Diaz, Y. (2013). The impact of universal coverage schemes in the developing world: a review of the existing evidence Universal Health Coverage Studies Series (UNICO) No. 25. Washington, DC: The World Bank. Grossman, M. (2000). The human capital model. In A. J. Anthony, J. Culyer, & J. P. Newhouse (Eds.). Handbook of health economics (Vol. 1A, pp. 347–408). Amsterdam: Elsevier.

GSO. (2008). Comprehensive Results÷ on the Household Living Standards ÷ ÷ Survey 2006 (K^et Qua S^o Li^eu T^ong Ho’p v^e Khao Sa´t Mu’c S^ong H _ Gia Ðı`nh Na˘m 2006). Hanoi Statistical Publishing House, General Statistics Office, Socialist Republic of Viet Nam. Guo, S., & Fraser, M. W. (2010). Propensity score analysis: Statistical methods and applications. California: Sage÷ Publications Inc. ÷ Health Insurance Law. (2008). (Lu^at Bao Hi^e m Y T^e). 25/2008/QH12. Heckman, J. J., Ichimura, H., & Todd, P. E. (1997). Matching as an econometric evaluation estimator: Evidence from evaluating a job training programme. The Review of Economic Studies, 64(4), 605–654. Hsiao, W. C., & Shaw, R. P. (Eds.) (2007). Social health insurance for developing nations. Washington, DC: The World Bank. Jalan, J., & Ravallion, M. (2003a). Does piped water reduce diarrhea for children in rural India. Journal of Econometrics, 112, 153–173. Jalan, J., & Ravallion, M. (2003b). Estimating the benefit incidence of an anti-poverty program by propensity-score matching. Journal of Business and Economic Statistics, 21(1), 19–30. Jowett, M., Deolalikar, A., & Martinsson, P. (2004). Health insurance and treatment seeking behaviour: Evidence from a low income country. Health Economics, 13, 845–857. Lagomarsino, G., Garabrant, A., Adyas, A., Muga, R., & Otoo, N. (2012). Moving towards universal health coverage: Health insurance reforms in nine developing countries in Africa and Asia. The Lancet, 380(8), 933–943. Lieberman, S. S., & Wagstaff, A. (2009). Health financing and delivery in Vietnam: Looking forward. Washington, DC: The World Bank. Madans, J. H., Loeb, M. E., & Altman, B. M. (2010). Measuring disability and monitoring the UN Convention on the Rights of Persons with Disabilities: The work of the Washington Group on Disability Statistics. BMC Public Health, 11(S4), 1–8. Mont, D., & Cuong, N. V. (2011). Disability and poverty in Vietnam. The World Bank Economic Review, 25(2), 323–359. Nguyen, C. V. (2012). The impact of voluntary health insurance on health care utilization and out-of-pocket payments: New evidence from Vietnam. Health Economics, 21, 946–966. Nhat, H. (2012). Using barcodes to reduce hospital overcrowding. (Du`ng ÷ ÷ ÷ ÷ ma˜ vach d^e giam tai b^enh vi^en). Tu^o i Tr Online. _ _ O’Donnell, O., van-Doorslaer, E., Wagstaff, A., & Lindelow, M. (2008). Analyzing health equity using household survey data: A guide to techniques and their implementation. Washington, DC: The World Bank. Rosenbaum, P., & Rubin, D. (1983). The central role of the propensity score in observational studies for causal effects. Biometrika, 70, 41–55. Rosenbaum, P. R., & Rubin, D. B. (1985). Constructing a control group using multivariate matched sampling methods that incorporate the propensity score. The American Statistician, 39(1), 33–38. Sepehri, A., Sarma, S., & Oguzoglu, U. (2011). Does the financial protection of health insurance vary across providers? Vietnam’s experience. Social Science & Medicine, 73, 559–567.

394

WORLD DEVELOPMENT

Somanathan, A., Huong, D. L., & Tran, T. V. (2013). Integrating the poor into universal health coverage in Vietnam Universal Health Coverage Studies Series (UNICO) No. 24. Washington, DC: The World Bank. Stewart, F. (2005). Groups and capabilities. Journal of Human Development, 6(2), 185–204. Tran, T. V., Hoang, P. T., Mathauer, I., Nguyen, P. K. T. (2011). A health financing review of Viet Nam with a focus on social health insurance. World Health Organization. Trujillo, A. J., Portillo, J. E., & Vernon, J. A. (2005). The impact of subsidized health insurance for the poor: Evaluating the Columbian experience using propensity score matching. International Journal of Health Care and Economics, 5, 211–239. United Nations. (2013). A new global partnership: Eradicating poverty and transform economies through sustainable development. The Report of the High-Level Panel of Eminent Persons on the Post-2015 Development Agenda. New York: United Nations Publications. Wagstaff, A. (2007). Health insurance for the poor: Initial impacts of Vietnam’s Health Care Fund for the poor. The World Bank. Wagstaff, A. (2010). Estimating health insurance impacts under unobserved heterogeneity: The case of Vietnam’s Health Care Fund for the Poor. Health Economics, 19, 189–208. Wagstaff, A. (2011). (Producer). Beyond universal coverage Part I. Let’s talk development. Retrieved from .

Wagstaff, A., & van-Doorslaer, E. (2003). Catastrophe and impoverishment in paying for health care: With applications to Vietnam 1993– 1998. Health Economics, 12, 921–934. Washington Group. (2008). The measurement of disability: recommendations for the 2010 round of censuses. WHO. (2010). The world health report health systems financing: The path to universal coverage. Geneva: World Health Organization. Xu, K., Evans, D. B., Carrin, G., Aguilar-Rivera, A. M., & Evans, P. M. T. (2007). Protecting households from catastrophic health spending. Health Affairs, 26, 972–983. Zweifel, P., & Manning, W. G. (2000). Moral hazard and consumer incentives in health care. In A. J. Culyer, & J. P. Newhouse (Eds.). Handbook of health economics (Vol. 1). Amsterdam: Elsevier Science.

APPENDIX A. SUPPLEMENTARY DATA Supplementary data associated with this article can be found, in the online version, at http://dx.doi.org/10.1016/ j.worlddev.2014.06.008.

Available online at www.sciencedirect.com

ScienceDirect